diff --git a/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs b/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
index be5f71b6a..5a983f617 100644
--- a/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_experience_collector.rs
@@ -974,21 +974,14 @@ impl GpuExperienceCollector {
         let exp_h_b2_f32 = stream.alloc_zeros::<f32>(alloc_episodes * network_dims.3)
             .map_err(|e| MLError::ModelError(format!("alloc exp_h_b2_f32: {e}")))?;
 
-        // GPU-native domain randomization kernels
-        let domain_rand_starts_kernel = {
-            let context = stream.context();
-            let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
-                .map_err(|e| MLError::ModelError(format!("dr cubin: {e}")))?;
-            module.load_function("domain_rand_episode_starts")
-                .map_err(|e| MLError::ModelError(format!("domain_rand_starts load: {e}")))?
-        };
-        let domain_rand_params_kernel = {
-            let context = stream.context();
-            let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
-                .map_err(|e| MLError::ModelError(format!("dr cubin2: {e}")))?;
-            module.load_function("domain_rand_sim_params")
-                .map_err(|e| MLError::ModelError(format!("domain_rand_params load: {e}")))?
-        };
+        // GPU-native domain randomization + saboteur kernels (ONE cubin load)
+        let exp_module_extra = stream.context()
+            .load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
+            .map_err(|e| MLError::ModelError(format!("exp cubin extra: {e}")))?;
+        let domain_rand_starts_kernel = exp_module_extra.load_function("domain_rand_episode_starts")
+            .map_err(|e| MLError::ModelError(format!("domain_rand_starts load: {e}")))?;
+        let domain_rand_params_kernel = exp_module_extra.load_function("domain_rand_sim_params")
+            .map_err(|e| MLError::ModelError(format!("domain_rand_params load: {e}")))?;
         let domain_rand_params_buf = stream.alloc_zeros::<f32>(alloc_episodes * 5)
             .map_err(|e| MLError::ModelError(format!("alloc dr_params: {e}")))?;
 
@@ -1004,21 +997,11 @@ impl GpuExperienceCollector {
             .map_err(|e| MLError::ModelError(format!("alloc saboteur_best: {e}")))?;
         let saboteur_best_return_buf = stream.alloc_zeros::<f32>(1)
             .map_err(|e| MLError::ModelError(format!("alloc saboteur_best_return: {e}")))?;
-        // Load saboteur kernels from experience cubin
-        let saboteur_generate_kernel = {
-            let context = stream.context();
-            let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
-                .map_err(|e| MLError::ModelError(format!("saboteur cubin: {e}")))?;
-            module.load_function("saboteur_generate_params")
-                .map_err(|e| MLError::ModelError(format!("saboteur_generate load: {e}")))?
-        };
-        let saboteur_select_kernel = {
-            let context = stream.context();
-            let module = context.load_cubin(EXPERIENCE_KERNELS_CUBIN.to_vec())
-                .map_err(|e| MLError::ModelError(format!("saboteur cubin2: {e}")))?;
-            module.load_function("saboteur_select_best")
-                .map_err(|e| MLError::ModelError(format!("saboteur_select load: {e}")))?
-        };
+        // Saboteur kernels from same module (no extra cubin load)
+        let saboteur_generate_kernel = exp_module_extra.load_function("saboteur_generate_params")
+            .map_err(|e| MLError::ModelError(format!("saboteur_generate load: {e}")))?;
+        let saboteur_select_kernel = exp_module_extra.load_function("saboteur_select_best")
+            .map_err(|e| MLError::ModelError(format!("saboteur_select load: {e}")))?;
 
         // #31 Load bottleneck tanh+concat kernel from utility cubin (if active)
         let bn_tanh_concat_fn = if bn_dim_from_params > 0 {
diff --git a/crates/ml/src/cuda_pipeline/mod.rs b/crates/ml/src/cuda_pipeline/mod.rs
index 2506132e4..15fc3e80d 100644
--- a/crates/ml/src/cuda_pipeline/mod.rs
+++ b/crates/ml/src/cuda_pipeline/mod.rs
@@ -1107,7 +1107,7 @@ mod tests {
         let cfg = PpoCollectorConfig::default();
         assert_eq!(cfg.n_episodes, 128);
         assert_eq!(cfg.timesteps_per_episode, 500);
-        assert_eq!(cfg.total_experiences(), 128_000); // 2x counterfactual
+        assert_eq!(cfg.total_experiences(), 64_000); // PPO: no counterfactual doubling
         assert!((cfg.gamma - 0.99).abs() < f32::EPSILON);
         assert!((cfg.gae_lambda - 0.95).abs() < f32::EPSILON);
     }